Skip to main content
. 2022 May 18;20:2484–2494. doi: 10.1016/j.csbj.2022.05.031

Fig. 3.

Fig. 3

The structure of explainable 3DU-Net. The modality of the input is set to 8, which is the maximum number of volumes a patient can have. If a patient has fewer than 8 DCE-MRI volumes, the positions with absent volumes are set to empty. The output is the tumor segment annotation. The last hidden layer of the encoder phase is the DLR features. Two explanation tools (Gradient map and Gradient*image map) are used to increase the explainability of the 3DU-net.